Answering Legal Questions with LLMs
Large language models are being explored as tools to answer legal questions, but their tendency to hallucinate, miss context across statutes and case law, and present guesses with unwarranted confidence makes unsupervised use risky. Commenters see promise in constrained roles—summarizing dense legal text, aiding research via RAG-style systems, or serving as an “intern” for lawyers—so long as a qualified human verifies the output and bears responsibility. There is broad skepticism that LLMs can safely replace legal professionals in complex, precedent-driven systems, though many expect them to reshape workflows, lower costs for routine matters, and intensify debates over access, liability, and professional standards.
Scope and Complexity of Law
- Many commenters doubt LLMs can correctly answer legal questions because real-world law is highly interconnected: statutes cross-reference, depend on regulations, guidance, case law, and local practice.
- GDPR and EU law are cited as examples where you must read statutes together with opinions, guidelines, and court decisions.
- Several argue that for genuinely hard legal questions, identifying which materials even matter is itself expert work; if a lawyer must curate all inputs, they might as well write the answer.
RAG, Context Windows, and Technical Limits
- There is extensive debate on Retrieval-Augmented Generation:
- Supporters say RAG is ideal for legal/medical use: constrain answers to uploaded documents, provide citations, and dramatically reduce hallucinations.
- Skeptics counter that RAG only helps; LLMs still hallucinate, especially at long context lengths or when instructions get diluted.
- Some call RAG “as safe as SQL,” others strongly dispute this, noting legal facts and reasoning don’t map neatly to product-catalog–style databases.
- Long-context models (e.g., Gemini 1.5) are praised for handling huge manuals, but people note quality still degrades with very large prompts.
Role of LLMs in Legal Practice
- Broad consensus: LLMs should be tools “in service of” lawyers, not standalone advisors to end users.
- Current high‑value use cases:
- Summarizing and simplifying legal language (with human review).
- Acting as a “junior associate” / research assistant / personal librarian to surface relevant passages and cases.
- Direct, unreviewed use for filings has already produced public failures (made‑up case law), reinforcing the need for human verification and professional liability concerns.
Trust, UX, and Product Design
- Many want systems that foreground sources and minimize “answer-y” prose, to discourage blind trust and reduce automation bias.
- Some suggest semantic/embedding search plus highlighted passages may be safer than full generative answers.
- Others highlight that human laziness is the unsolved problem: even with citations, many users won’t rigorously check.
Democratization vs. Inequality and Systemic Effects
- One camp expects LLMs to democratize access to legal information, enabling cheap complaints, Q&A, translation, and assistance for less powerful parties.
- Another worries laws will grow more complex once elites can offload complexity to AI, further advantaging those with the best tools and deepening opacity.
- There is discussion of potential arms races (dueling legal AIs) and protectionism by legal institutions.
Comparisons to Other Professions
- Programming is widely seen as an area where LLMs already add clear value; that success fuels optimism for law and medicine.
- Others draw analogies to self-driving cars: getting from “pretty good” to the reliability needed in safety‑critical domains may be a very long tail.